I have had to think carefully about how technical results are phrased, especially when a sentence gets rewritten to sound cleaner. One common mistake is turning a 20% reduction in error into a 20% improvement in accuracy. Because a 20% cut in error shrinks a small remaining number, accuracy itself can barely move. The number survives the rewrite, but what it describes changes completely, and the error can slip past a quick read.
Evidence Desk is a browser tool I built with Codex to catch errors like this. You supply a draft and its source files. The tool splits the draft into claims, finds the closest passage in the sources for each one, and runs explicit numeric, metric and causal checks. Each claim is labeled Supported, Potential mismatch, or Insufficient evidence. Codex helped implement the interface, checking logic and tests, while I decided what the tool should check and reviewed its behavior.
Checking is a different job from writing. It asks whether the evidence actually says what the claim says, which is narrower than asking whether the claim is true. Instead of using a second generative model as reviewer, Evidence Desk attaches every warning to the passage it relies on and a stated reason, so the writer has something specific to push back on.
The pipeline has limits. Claim extraction is crude, and retrieval is plain lexical overlap with hand-set thresholds. A high match score only shows how closely a sentence was derived from its source, and a distorted rewrite is a close derivation by definition. In the built-in sample, the tool found the right passage but still missed the mismatch, which shows where rule-based checks break down.
Source: Towards Data Science · Summarized by HeadlinesBriefing